This dataset contains 10 manually constructed blind spot cases designed to evaluate reasoning, instruction following, and factual robustness of the model:

CohereLabs/tiny-aya-water

The goal of this dataset is to identify systematic weaknesses and failure modes of the model under controlled experimental conditions.

Dataset link

Hugging Face Dataset

Dataset characteristics

• Small evaluation dataset (10 cases)
• Designed for LLM robustness testing
• Focus on reasoning failures and blind spots
• Useful for controlled experimental analysis

Potential use cases

• Evaluation of large language models
• Testing instruction-following robustness
• Studying failure modes in generative AI systems